Use LMDB to store vectors in PAV backend

[osma]

Similar to #363; related to #339

When the PAV backend is trained, it sends all documents through the source projects and aggregates their suggestion vectors in memory. This can take up a significant amount of RAM - not quite as much as with nn_ensemble but still several GB in typical YSO settings.

If we instead streamed the vectors to a LMDB database, and then read them back from the LMDB in batches, the backend could scale to much larger training data sets. An additional benefit would be that the LMDB could be retained on disk, so that another training run could be made using the same documents but different hyperparameters (with the --cached option - see #342), without having to process the documents again, so it would be much faster.

LMDB seems to be ideal for this as it is very fast and supports streaming style operations both for reading and writing. It will introduce an additional dependency, though.

[osma]

After switching to sparse vectors (#379) in the PAV backend the RAM usage is now much lower, so this is not so crucial anymore.